Telemetry becomes part of the agent platform
Google Cloud has expanded its AI agent observability and governance capabilities across BigQuery and Looker, giving platform teams one telemetry model for first-party data agents and custom agents built with frameworks such as Agent Development Kit or LangGraph. The design captures operational data, lands selected records in customer-owned BigQuery tables and exposes dashboards in the tools data teams already use. Google is targeting the production gap between deploying an agent and understanding how it behaves, what it touches and what it costs.
First-party Conversational Analytics agents can emit Cloud Monitoring metrics, Cloud Trace spans and BigQuery conversation logs without a separate export pipeline. Custom agents can stream compatible structured telemetry through SDK instrumentation. That common model matters when organisations have agents from several teams and frameworks. A central operations group can compare sessions, latency, sources and tool behaviour without forcing every developer to invent a logging schema or send sensitive records into an unrelated monitoring service.
Conversation logging is now generally available for BigQuery-based agents and writes to tables owned by the customer. BigQuery Studio dashboards track usage, active sessions, frequently used knowledge sources and latency. Looker System Activity dashboards, also generally available, report engagement and token consumption, while response ratings and query review remain in preview. Open-source templates and a Looker Block extend the same records into custom reporting for mixed first-party and third-party agent estates.
Privacy changes what teams should record
The most important design choice is the separation between aggregate operations data and conversation content. Metrics and sanitised first-party traces exclude prompts, responses, internal thoughts and generated SQL. Teams can opt into deeper per-agent conversation logging when investigation requires transcripts. This avoids making the most sensitive content a default by-product of monitoring, but organisations still need retention rules, access reviews and a clear statement of who may inspect a user’s conversation and for what purpose.
Google says logging fails closed when a knowledge source uses row-level security, column policy tags or dynamic masking. If the safety check detects those controls, it skips conversation logging for that turn while allowing the user session to continue. That balance protects governed data without making a logging failure an availability failure. Teams should test it explicitly, because a silent absence in telemetry can otherwise be mistaken for an agent that never accessed sensitive material.
Custom ADK and LangGraph agents can record fuller tool-execution trees and multimodal payloads through the BigQuery Agent Analytics plugin. More detailed traces improve root-cause analysis, yet they also widen the data boundary. Tool arguments may contain credentials, customer records or proprietary inputs that the first-party sanitised path intentionally removes. Security teams should define field-level redaction and test representative failures before enabling complete trace capture across a production agent fleet.
FinOps controls move closer to each turn
The release also adds more granular cost visibility. Each logged turn can record input, output, thought, cached and total tokens alongside latency. BigQuery byte-scan limits can block an oversized query using a dry-run estimate, while preview token quotas can be applied per project, user or agent. These are useful controls because an autonomous workflow can multiply cost through retries, broad queries and long reasoning loops even when every individual model call appears affordable.
Cost dashboards should be tied to acceptable outcomes rather than token minimisation alone. A more expensive agent may be worthwhile if it resolves a case correctly and avoids human rework. Conversely, a cheap agent that repeatedly queries the warehouse can create hidden operational expense. Platform teams should connect telemetry to business identifiers, without exposing unnecessary personal data, so they can compare spend, latency, correction rate and completed work across versions.
Google is pairing runtime monitoring with Prism, an open-source evaluation tool for Conversational Analytics agents grounded in BigQuery or Looker. Test suites can use reusable prompts and assertions, then stream results into partitioned BigQuery tables beside production logs. First-party BigQuery evaluations are also available in preview. This creates a practical release gate: teams can compare prompts, models and schema changes before a new version reaches users, then investigate whether production behaviour matches the benchmark.
Evaluation closes the deployment loop
The broader lesson is that enterprise agents need an operating system for evidence, not just a model endpoint. Google’s approach makes BigQuery both a data source and a governed record of agent activity, with Looker serving administrators and business reviewers. Australian organisations should verify regional logging options and whether their selected telemetry path meets local privacy, contractual and records-management duties. They should also budget for the storage, retention and analysis generated by detailed traces across every production environment before enabling it broadly. That consolidation will be valuable if customers maintain strict access, retention and evaluation discipline. Observability should make an agent more accountable without turning every conversation into an indefinitely stored surveillance record.